On Mon, Jun 21, 2010 at 11:14:16PM -0700, Andrew Morton wrote:
> On Tue, 22 Jun 2010 15:44:09 +1000 Dave Chinner <david@xxxxxxxxxxxxx> wrote:
>
> > > > And so on. This isn't necessarily bad - we'll throttle for longer
> > > > than we strictly need to - but the cumulative counter resolution
> > > > error gets worse as the number of CPUs doing IO completion grows.
> > > > Worst case ends up at for (num cpus * 31) + 1 pages of writeback for
> > > > just the first waiter. For an arbitrary FIFO queue of depth d, the
> > > > worst case is more like d * (num cpus * 31 + 1).
> > > Hmm, I don't see how the error would depend on the FIFO depth.
> >
> > It's the cumulative error that depends on the FIFO depth, not the
> > error seen by a single waiter.
>
> Could use the below to basically eliminate the inaccuracies.
>
> Obviously things might get a bit expensive in certain threshold cases
> but with some hysteresis that should be manageable.
That seems a lot more... unpredictable than modifying the accounting
to avoid cumulative errors.
> + /* Check to see if rough count will be sufficient for comparison */
> + if (abs(count - rhs) > (percpu_counter_batch*num_online_cpus())) {
Also, that's a big margin when we are doing equality matches for
every page IO completion. If we a large CPU count machine where
per-cpu counters actually improve performance (say 16p) then we're
going to be hitting the slow path for the last 512 pages of every
waiter. Hence I think the counter sum is compared too often to scale
with this method of comparison.
Cheers,
Dave.
--
Dave Chinner
david@xxxxxxxxxxxxx
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
